Build software, orchestrating from a CLI - Problem Engineering Part 3
Target Audience: Anyone who develops software, tests it, or is into DevOps.
In my last post, we explored how to break free from PPP (Prompt, Paste, Pray) by acting as a "Problem Engineer." We used specialized agents (ChatPRD, v0, Supabase) to build a full-stack app, showing that engineering constraints are more effective than writing isolated features.
But there was a catch: You were still the coordinator. You were switching tools, copying context, and manually driving the flow
Let's take the next step. We are moving from manual agent handoffs to central orchestration. We will use one central tool to run multiple agents, building a React frontend and Python backend, entirely from the command line. We will then use GitHub issues and PRs to enhance the software, mirroring what you do in your daily work.
Here is how you operationalise Problem Engineering - at the orchestration layer.
The Insight: Mapping the "Goldilocks Zone" - to Code
To orchestrate agents effectively, we must map our theoretical framework directly into the environment:
Note: I am using claude but this can easily be codex, or copilot, or aider, or other tools, some of which I showed on this graphic.
We are establishing a contract-first pipeline for new projects, existing codebases, and continuous improvement. This is not to establish best practice, or even near what I would call great quality, but good as a next step for developers on the AI journey. More will follow...
Let’s get our toes wet - The Walkthrough: The "Stranger Things" Calculator
We will build a themed calculator using a React/Next.js SPA (Port 3004) and a Python FastAPI backend (Port 8004), containerised with Docker. I choose Stranger Things, you may go with Bridgerton or Spider-Man, or whatever makes you tick.
Phase 1: The Setup
Instead of bouncing between browser tabs, we set up our environment to govern the agents for us:
# Initialize: Create an empty GitHub repo and install Claude Code
git clone https://github.com/vickybiswas/agent-demo
cd agent-demo
curl -fsSL https://claude.ai/install.sh | bash
# Equip Plugins: Install specialized knowledge (I suggest frontend-design, pr-review-toolkit, security-guidance)
claude plugin install @anthropic/[plugin-name]
# Bind Tools (MCP): Give the agent access to real tools:
# playwright for UI testing
npm install @playwright/mcp && claude mcp add playwright npx @playwright/mcp@latest
# code-review-graph for architecture review
claude plugin marketplace add tirth8205/code-review-graph && claude plugin install code-review-graph@code-review-graph
# github to read/write PRs directly
claude mcp add-json github '{"type":"http","url":"https://api.githubcopilot.com/mcp","headers":{"Authorization":"Bearer '"$(grep GITHUB_PAT .env | cut -d '=' -f2)"'"}}'
I orchestrated a skill that would help set up hooks, Five Agents, Three Validators (nextjs, fastapi-validator, docker-validator), and Orchestration Files (like CLAUDE.md, REGRESSION.md, CREATE.md, STARTUP.md). In real life, these should be handmade and revised by the team periodically.
I am using Claude code pro for this demo; you can install and use it for free(see Appendix for instructions). However, it is important to understand that the free versions limit context and speed based on the models you choose. For this article, I used Haiku with Pro in yolo mode -> claude --dangerously-skip-permissions --model haiku
Note: Ensure docker is running so claude can use it for testing
Tell Claude: Use the repo-setup skill and INSTRUCTIONS.md to set up Agents, skills, hooks, all CLAUDE.md, etc. for agentic development.
Phase 2: Execution (The New Project)
With constraints locked, we issue high-level instructions:
Tell Claude: Use INSTRUCTIONS.md, the 3 Claude.md files, skills, agents, etc., to build/test a Stranger Things themed calculator with a animated and interactive UI. Use subagents to parallize work.
This will take roughly 20 minutes to build and run the entire code, test cases, documentation, etc. If it asks, answer to your best knowledge, or just say 'follow instructions given'. Once it finishes you will see something like:
Recommended by LinkedIn
Fire it up.
docker compose up
You can access the frontend and Backend (swagger).
Phase 3: Day 2 Operations (Established Projects)
Problem Engineering isn't just for greenfield apps. Because our constraints are defined, we can instruct the agent to orchestrate updates safely. Let's use GitHub Issues to initiate new features and changes, do a prelim RCA, make the fixes, raise a PR:
GitHub Issues: Update Calculator - Update the calculator to a Scientific one. Also, ensure you add elements, framer animations, interactions, colors, and sounds to change UI to a new Spiderman based theme.
Now let's ask Claude to get this done for us:
Tell Claude: Using fix-github-issue implement issue [issue id or url] and test it E2E thoroughly.
You should soon see something like:
The Takeaway
We only covered a frontend and backend here, but the possibilities are endless. Understanding the Problem Engineering is critical here covering concisely:
Are you ready to stop typing and start orchestrating? Try it yourself and let me know how "Problem Engineering" changes your workflow.
Pro-Tip: Use git worktrees to run multiple Claude sessions in parallel, acting as an Agent Team, or explore the agent team feature.
Appendix
1. Claude Code for free (Severely dependent on model)
2. Time and Cost
3. Quality of files and prompts
The prompts and files used here are shortcuts taken for demo purposes and are not anything near to what I would use in production. In coming posts I will approach on that as well. To start I would want the md files to be between 100 and 200 lines.
4. Ask claude for status if it waits too long
Lovely. Exact and precise.